Copilot Studio is the default answer for a lot of companies that run on Microsoft 365. It's already in the tenant, it's sold by the same account team, and it promises agents without engineers.
But most of the complaints we hear about it, and most of the limits we run into ourselves, come back to one thing. Copilot Studio isn't a platform designed for agents. It's the Power Platform, the same machinery behind Power Automate, Power Apps and the old Power Virtual Agents, with an agent builder put on top. The licensing, the connectors, the environments, the testing and the channels all carry that inheritance, and an agent needs something different from each of them.
In short:
- Search built before AI search. Answers draw on the top three SharePoint results, and large files aren't read.
- The meter runs before the agent does. You pay for seats and capacity up front, and agents stop when capacity runs out.
- Teams first, everything else by hand. Other channels need manual sign-in, and Teams channels can't use SharePoint knowledge.
- Outside the connector catalogue, you're on your own. Custom connections mean Premium licenses, MCP setup problems and workarounds.
- You can't see why it did what it did. Transcripts are short-lived and built-in testing misses wrong answers.
- You're the test subject. Apps, names and models change or retire on Microsoft's schedule.
- One model menu, one harness. Microsoft picks the models, runs the orchestration and upgrades the default for you.
- The Power Platform mold. Environments, solutions and Microsoft-only terms, with connections that break when people leave.
Search built before AI search
Most agents are only as good as what they can find: the right file, the right person, the right task. In Copilot Studio, finding things in SharePoint runs through Microsoft Graph, on top of Microsoft Search. Microsoft has added a semantic index for Copilot, but the foundation is an enterprise search built before language models, and in our view it shows.
Microsoft's own troubleshooting page is candid about the limits:
When Copilot Studio searches SharePoint, it uses only the top three search results to summarize and generate a response.
- Three results. If the right document isn't in the top three, the agent doesn't see it.
- File size. Without a Microsoft 365 Copilot license in the tenant, generative answers only process files up to 7 MB. Larger files are found by the search and then not read.
- Freshness. SharePoint knowledge is synced periodically, and new or changed files can take hours to show up in answers.
Newer approaches to retrieval let the model search again when the first results are wrong: refine the query, open the file, follow a reference. Copilot Studio's approach is closer to one search and a summary. The community forums are full of the result: documents the agent can see in the builder but not once it's published, and agents that say they can't find anything in the files they were given.
The meter runs before the agent does
Copilot Studio is paid for in more than one way at once. There are seats for the people who build and use agents, and there are Copilot Credits for what the agents do: a capacity pack of 25,000 credits a month for $200, or a cent a credit on pay-as-you-go, according to Microsoft's licensing guide. Capacity is enforced monthly and unused credits don't carry over, so a company can be paying for AI creation before a single useful agent exists.
What an agent costs is also hard to predict, because credits are charged per action, not per conversation. A scripted answer is cheap. A generative answer costs more, grounding in the company's own data more again, and a reasoning model or voice far more. The same agent can cost very different amounts depending on how people use it.
And the agent only runs while the meter does. Licensing guides describe agents being blocked, mid-conversation, once a tenant passes 125% of its prepaid capacity, and agent flows stopping as soon as capacity runs out. No capacity, no agent: what you built is a setting in your tenant, and it switches off when the bill stops. We go through licensing choices in more depth in The AI Licensing Playbook.
Teams first, everything else by hand
Copilot Studio works most smoothly in Teams and Microsoft 365 Copilot, where sign-in is automatic. Every other channel needs authentication set up manually, and Microsoft's own pages disagree on which channels, like Slack and WhatsApp, are available at all.
Even inside Teams, the fit is narrower than it looks. According to Microsoft's Teams documentation, agents in group chats and channels can't use knowledge sources that need the user to sign in, which includes SharePoint. An agent that answers well in a one-to-one chat can go quiet in the team channel where people actually work.
Outside the connector catalogue, you're on your own
Inside Microsoft's catalogue of connectors, connecting an agent to a system is easy. Outside it, things get hard quickly. Certified custom connectors are only released as Premium. Connecting your own systems through the Model Context Protocol (MCP) is supported, but the Microsoft Q&A forums show how often it goes wrong: servers rejected by tenant data policies after working before, failed setups that leave half-created connectors behind, sign-in methods that aren't supported, and conflicting answers about which license is needed.
Even something as ordinary as producing a PDF is a licensing decision. Copilot Studio's own document output produces Word files from a template. Turning one into a PDF means the Premium Word Online (Business) connector, which won't convert files labelled Confidential or Highly Confidential, or Adobe, a third-party connector, or a workaround through OneDrive. For a platform that sells itself on connecting to everything, the edges come quickly.
You can't see why it did what it did
When an agent gives a wrong answer, the first question is why. Copilot Studio's activity map helps for simple cases, but the tools around it are thin. Conversation transcripts are deleted after 30 days by default. Errors often arrive without a reason. Microsoft's own documentation says its built-in evaluations don't test for gaps in what the agent knows, and that passing them doesn't guarantee an appropriate answer.
One consultancy that tested its agents like production software found the gap the hard way:
The first time we asked our live Copilot agents the questions we'd written for them, six of twenty-three came back wrong.
Their deployment checks had passed. The parts were there; the answers weren't right. For anything regulated, or anything where a wrong answer costs money, an agent you can't trace and can't reliably test is hard to defend.
You're the test subject
Companies often choose Microsoft for stability. Copilot Studio moves fast enough that it rarely feels stable:
- The Teams app is going away. Microsoft is retiring the Copilot Studio app in Teams: editing ends in January 2027 and the app is removed in February 2027.
- Names keep changing. Power Virtual Agents became Copilot Studio. Agent Builder became "Copilot Studio lite", then Agent Builder again.
- Models are retired for you. Agents on Claude Sonnet 4.5 were moved to Sonnet 4.6 automatically unless an admin opted out in time, and Microsoft has announced the end of support for other models.
Features ship in preview, change, and get shelved if they don't catch on. That's a hard fit for a product sold as the safe enterprise choice, where what you build today may have to be rebuilt when the platform moves.
One model menu, one harness
Copilot Studio's model list is wider than it was. As of Microsoft's October 2026 update, GPT-5.5 Chat is the default and Claude Opus 4.8 is generally available. But it's still Microsoft's menu. There's no supported way to add your own model as an agent's main model. Models from Microsoft Foundry can be brought in only for individual prompts, not to run the agent, and Microsoft's documentation lists the GPT-5 family as unsupported even there.
The harness, the orchestration that decides which tool to call and when, isn't a choice at all. And the model under it changes on Microsoft's schedule: when Copilot Studio upgrades its default, you can keep your agent on its current model for up to 30 days. After that, your agent's behavior changes whether or not you changed anything.
The Power Platform mold
Building in Copilot Studio feels like building in the rest of the Power Platform, because it is. Agents live in environments. Makers need the Environment Maker role. Moving an agent from test to production goes through solutions and pipelines. Agents are made of topics, agent flows and connection references, and they're shared and governed through Dataverse.
None of those terms mean anything outside Microsoft. Elsewhere, agents are built from things that travel: MCP servers, tool definitions written to an open standard, code in a repository, evaluations that run against any model. What a team learns building in Copilot Studio stays in Copilot Studio, and so does what it builds.
The same inheritance shows when people leave. An agent can outlive its owner, but its connections often don't: tools signed in with the maker's credentials, flows the maker owned, and knowledge sources only they could reach. Ownership can be reassigned, but the new owner needs a Copilot Studio license, and reassigning it doesn't repair any of that.
When it's still the right call
None of this means Copilot Studio is never the answer. We'd use it when:
- Everything lives in Microsoft. The data and actions stay in Outlook, SharePoint, Teams and Office, and nothing needs to connect outside the catalogue.
- The workflows are simple and built around email and documents. Triage and draft replies, summarize a folder, fill a template.
- It's one person talking to the agent. In a one-to-one chat in Teams or Outlook, not a group chat or a public channel.
- Your SharePoint is in decent shape. Sensible permissions, current files and few confidential labels, because the agent inherits whatever is there.
Once a workflow needs a system outside Microsoft, a model of your choosing, real volume, or a record of why the agent did what it did, you're paying for the costume and fighting the platform underneath it. That's the point where we'd build the agent in code instead: connections through MCP, the model that fits the job, and tests and logs the business owns.